[Submitted on 15 Dec 2025 (v1), last revised 28 Jul 2026 (this version, v2)]

View PDF HTML (experimental)

Abstract:Tile-based many-Processing Element (PE) accelerators can achieve competitive performance on General Matrix Multiplication (GEMM), but they are extremely hard to program, as their optimal software mapping is deeply coupled with hardware design which is unwieldy to manual deployment. We propose "Design in Tiles (DiT)", an automated framework connecting a deployment toolchain with a configurable executable model for these accelerators. For evaluation, we apply our framework to GEMM targeting a large acceleration configuration (e.g., 32x32 tiles, 1979 TFLOPS@FP8, 4 TB/s Bandwidth) comparable to an NVIDIA GH200. We achieve higher PE utilization than GH200 with its expert-tuned GEMM libraries, achieving 1.2-2.0x speedup across diverse matrix shapes.

Submission history

From: Aofeng Shen [view email]
[v1] Mon, 15 Dec 2025 18:33:04 UTC (953 KB)
[v2] Tue, 28 Jul 2026 16:54:19 UTC (307 KB)